For regulated mid-sized companies · Stuttgart, Germany

AI does the groundwork.
Your team decides.

We build software for your business processes. It records who decided what, and on what basis. Any later change to that record can be detected.

The software shown here is already running. Try the demo. Your own copy of Approval, deleted after 24 hours. No sign-up.

What this is about

Before a team works with AI, two questions need an answer: which data may a model receive, and who checks what it produces? Our libraries enforce the data rules and record every review decision. Here is how, by example.

An example

Ms Berger and Kai

Ms Berger keeps the customer records in accounts. Kai is her team’s AI assistant: he reads incoming requests and drafts the changes. Whether they are made is Ms Berger’s call.

Data disclosure and missing decision records

  • Data sent outside your organisation

    Confidential content is sent to an outside AI service without anyone checking it first.

  • Decisions without proof

    Without a decision record, it is hard to reconstruct later who approved a change and why.

Every request, the same path

The path of an AI request

Each station passes on only what it has checked. That is how a request becomes a decision you can prove.

ExampleA customer changes their bank detailsInvented case, for illustration.

  1. 01

    Request

    A department sends in a case, with its documents.

    The email: “Please change my bank details. New IBAN attached.”

  2. 02

    Classification

    Public, internal or confidential — checked before every model call. Confidential content stays in your environment.

    Name and IBAN are confidential, so Kai uses a model in your environment. These details stay there.

    disclosurepii

  3. 03

    AI proposal

    The AI prepares a proposal with the current values beside it. It writes nothing to the database itself.

    Kai drafts the change: new IBAN in the customer record, the old one beside it, the email as the source.

    proposal

  4. 04

    Human review

    A responsible person accepts, edits or rejects the proposal, in part if need be.

    Ms Berger checks the proposal against the email and accepts it.

    agent-ui

  5. 05

    Ledger

    Each applied change is chained into the ledger: who, when, on whose behalf. Later edits show.

    Entry: bank details changed, accepted by Ms Berger, on the customer’s behalf, 14 Oct, 9:42.

    agent-ledger

  6. 06

    Evidence

    The evidence for a period is generated from the ledger, ready for the audit.

    At the next audit: every change of bank details this quarter, and who accepted each one.

    evidence

Why octabits

Libraries for classification, review and decision records

  • Rules your IT can read

    Every station is an open-source package, 29 in all. From November 2026 your IT or an outside auditor can read what they do.

  • We use them ourselves

    Approval, Reynt and Freiday run on the same libraries.

  • You talk to the architect

    Daniel Hartmann is a software architect with 25 years of experience. He is your contact at octabits and develops your application.

A real product: Approval

Software approval, with the AI doing the groundwork

Every new piece of software needs approval from data protection and information security. That means reading the documents, collecting the facts, writing the assessments. In Approval the AI does this groundwork — only your officers may sign.

The question on the table
“May we use this new software tool?”

A department requests software. Approval reads five vendor documents, drafts both assessments and puts them in front of your officers.

  1. Facts extracted for your review

    Five documents, fourteen facts, each with its quote. A quote the document does not contain is rejected.

  2. Your team decides

    The model proposes and drafts. A person accepts each fact, and only the officers can sign.

  3. The evidence is recorded

    The log records every request, document version, decision and signature. Hash-linked entries make later changes detectable.

  4. It notices when the facts change

    A changed document invalidates the facts it no longer supports and reopens the approval based on them.

Watch it run

Approval first checks whether the software is already approved for the same use; if it is, that approval and its conditions apply and no new review starts. Otherwise it stores the vendor documents, and a new vendor’s sources and identity are checked first. A model extracts facts with supporting quotes; a quote that is not in the documents is rejected, and a person accepts, corrects or rejects each fact before the model drafts two assessments. The data protection and security officers each review and sign their part; one refusal rejects the request. The request, every document version, every decision and every signature are recorded in a tamper-evident log. When a vendor document changes, the affected fact is invalidated, the approval is out of date and a re-assessment supersedes it; an approval nearing its end is renewed from the current facts.

This shows the process used in Approval, with fictional requests, documents and figures.

We can implement this division of work for your processes, in a new application or your existing software.

A quotation that is not in the document is rejected at once.

Already running

Software we build and run ourselves

Freiday

Share when you are free, not what you are doing

Freiday shares your availability by link: a week view of busy and free times from all your calendars. The link expires on its own. Freiday stores no titles, places or attendees, only when you are busy.

about 10 h
To the first version in production
14
Described features, each with tested criteria
12
Automatic checks on every change

Reynt

Booking platform for holiday rentals

Reynt helps holiday rental operators manage bookings, offers, payments and invoices, with storefronts and compliance features. We develop and operate it using our published libraries.

11 months
From empty repository to production
19
Platform packages it runs on
EU
Developed and hosted

Working together

How it starts

  1. 01

    Describe your process

    How does the process work today, and who decides? A spreadsheet from your daily work is a good start. The first call is free.

  2. 02

    We build

    For the agreed scope: a first version in week one, delivery in 4–6 weeks

  3. 03

    You run it, we look after it

    We develop your application for a fixed price and agree a monthly fee for ongoing maintenance.

Development at a fixed price, with monthly maintenance. Prices in detail →

For management

What management gets

  • Runs on your infrastructureYou run the application on your own servers. We do not host it or process your data.
  • Belongs to youThe code is yours. No licence fees.
  • No per-user feesAdd as many users as you need at no extra charge.
  • Decisions on recordThe log records who approved each AI action and when. Later changes to the log are detectable.
Half a day
for Approval’s first version
29
open-source packages, MIT
0
signatures by a model
None
sub-processors of ours
Reynt
in production on the same parts

Your team reviews the AI’s proposals, and each decision is recorded. Confidential content stays in your environment.

Tell us briefly what you are working on and where you need help. Email is the fastest way to reach us; the person who would work with you will reply.